Table of Contents

• Coverage
• Dataset Description
  • Methods & Sampling
  • Data Processing Description
  • Problem Description
• Related Publications
• Related Datasets
• Parameters
• Instruments
• Project Information
• Funding

Acronyms: 

DESAFÍO = DisEntangling Seasonality of Active Flux in the Ocean (expedition name)
FoSI = Focused Shadow Imaging
LODP = Low-Optical-Density Particles
HODP = High-Optical-Density Particles
OD = Optical Density
ESD = equivalent spherical diameter
AZ = Azores
CI = Canary Islands
OO = Open Ocean
RGF = Relative Gel Fraction
ART ANOVA = Aligned Rank Transform (ART) Analysis of Variance (ANOVA)
ANCOVA = Analysis of Covariance

Particle data were collected during the DESAFÍO expedition (DisEntangling Seasonality of Active Flux in the Ocean; PID2020-118118RB-100) in the northeastern subtropical Atlantic Ocean, from 31 January to 2 March 2023. Sampling spanned a transect between the Azores (AZ; 41°5'32.39"N, 24°58'41.00"W) and the Canary Islands (CI; 28°41'60.00"N, 13°12'2.40"W). Twenty-two CTD/rosette casts were grouped into three geographic regions — CI (casts 1–6), the open-ocean station OO (casts 7–14, west of CI and south of the Azores), and AZ (casts 15–22) — and particle and hydrographic data within each cast were further partitioned into three depth strata: epipelagic (≤200 m), mesopelagic (200–1,000 m), and bathypelagic (>1,000 m). Four casts (5, 6, 18, and 20) have incomplete particle-image profiles due to unexpected camera shutdowns, likely caused by low battery voltage.

Particle images were processed following the pipeline of Huang and Bochdansky (2025), with analysis restricted to particles ≥23.5 μm (linear dimension) to exclude single-pixel noise. Two particle pools were defined by optical density (OD) threshold: low-OD particles (LODP; threshold 11, OD = 0.1466), which include the numerically dominant gel particles, and high-OD particles (HODP; threshold 29, OD = 0.2246), representing more opaque particles comparable to those detected by other optical instruments. Morphological metrics (equivalent spherical diameter, area, perimeter, aspect ratio, circularity, and OD) were extracted using MATLAB's regionprops(), with OD calibrated from brightness via an ensemble model. Both particle pools were resolved to 1-m depth bins for downstream analysis, including Theil–Sen regression with Mann–Kendall trend testing, Aligned Rank Transform (ART) ANOVA to assess regional and depth effects on abundance and total particle volume, Kolmogorov–Smirnov and Wilcoxon rank-sum tests for distributional/median comparisons, Wilcoxon signed-rank tests for paired day–night comparisons, calculation of the Relative Gel Fraction (RGF) from LODP and HODP abundances, and particle number spectrum analysis with ANCOVA to test slope homogeneity across regions and depths.

Huang, H., & Bochdansky, A. B. (2025). Optimizing an image analysis protocol for ocean particles in focused shadowgraph imaging systems. Frontiers in Marine Science, 12, 1539828.

NSF Award Abstract

Globally, the ocean removes more carbon dioxide than it releases into the atmosphere storing a portion of the excess carbon in the deep sea. Sinking particles, both living plankton and non-living detritus, are major contributors to this flux of carbon. Modern camera systems and image analysis techniques have made it possible to count, measure and classify these particles, thus providing oceanographers with a tool to estimate carbon transfers to the deep ocean at high resolution in space and time. Unfortunately, it is not enough to know the sizes of particles to estimate how fast these particles sink because shape and particle density also influence the sinking velocity. This project examines the velocities of individual particles as they sink into the deep ocean using a camera attached to a particle trap. For each of these particles, classification criteria, such as size, shape factors, optical density, and in the case of plankton, taxonomic identification, is determined and compared to their individual sinking velocities. This information serves to calculate overall sinking velocities from surveys of particles in the water column and thereby produce more reliable estimates of carbon fluxes from camera images. This project supports technology development in underwater imaging systems, graduate and undergraduate student education, and science literacy initiatives for middle-school students and their mentors through public outreach programs.

Shipboard and autonomous vehicle surveys of oceanic particle inventories hold great promise for estimating carbon fluxes at high temporal and spatial resolutions. However, while the sinking velocities of larger particles such as foraminifera shells and fecal pellets of salps, krill, and larger copepods are relatively well constrained, the dynamics of the smaller particle size pool (50–500 micrometers) remain more elusive. Despite their size and presumed slow sinking velocities, small particles occur in large numbers in the mesopelagic layer and sediment-trap material. Their abundance in the mesopelagic could be the result of deep mixing, or small particles could be remnants of digested larger particles, particles with a high excess density such as lithogenic dust particles, minipellets egested by protists, protist spores, or the result of fragmentation at depth due to the activity of flux feeders, among other possibilities. This project addresses some unanswered questions about the small particle pool by linking individually-resolved optical features with sinking velocities. Using Stokes’ law, excess density is being estimated from size and sinking velocity and then assigned to particles from optical surveys. A horizontally installed camera system records sinking velocities, sizes, and features of particles in a sediment trap attached to the Oceanic Flux Program mooring array. The recorded particles are being characterized using 1) classic image analysis, taking various shape factors into account; 2) opacity of individual particles; and 3) image classification with supervised and unsupervised deep learning using convolutional neural networks. A second identical camera surveys the particle inventory at the same station and time in the water column to integrate flux estimates over the existing and undisturbed particle pool. Niskin bottle samples and microscopic examination of particles augment the interpretation of image data. The results of this project contribute to the overarching goal of achieving higher predictive power for carbon flux models based on optical particle surveys.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.